What the model sees
How do the engine and the model share the work?
The agent is a hybrid. A deterministic engine reads the floor, lists every legal action with its exact cost, and annotates each action with what it unlocks and what it costs to use. The language model reads this list, reasons, and answers with the number of one action. The engine executes it. The model never computes a path or a fight, and it cannot name an action that does not exist.
Below is the real input at the first decision, as Qwen3.5-397B received it in a winning episode, with its reasoning and answer. The highlighted action is the one it chose.
- Damage and HP after
- exact HP cost of every fight and the HP that remains
- Status tags
- SAFE, MODERATE, RISKY, LETHAL, WINNABLE at ATK ≥ x, FREE if DEF ≥ x, DEAD END, alternative route
- Cost at maximum stats
- what the fight would cost after every remaining gem
- Lookahead tree
- what the action unlocks, up to three gateways deep
- Lookahead value, effective cost
- value of what is unlocked, net of the gateway fights
- Net stat-route line
- cost of reaching the gems behind an action and what they save
- Fight-order line
- HP wasted by taking a fight now instead of after a stat route
- Key lines and key budget
- keys used and gained, doors exposed; keys versus closed doors
- Boss line and HP budget
- whether the boss is winnable now and the HP the way to it needs
The engine also assists: it collects items and keys and takes fights that cost 0 HP without asking. A mechanical overridereplaces only choices that cannot be executed usefully: a door that leads nowhere, an unwinnable or a lethal fight. Every override is logged.